AI & Data Science Periodic Tables: How They Work Together

Aaron Baughman and Martin Keen introduce a unified perspective on Artificial Intelligence (AI) and Data Science, utilizing “periodic tables” to demystify the extensive jargon in both domains. They assert that while AI is currently prominent, its efficacy is deeply rooted in foundational data science principles. Effective AI models rely on meticulously prepared and cleaned data, and reciprocally, AI is now leveraged within data science for tasks such as data labeling and generating synthetic data, thereby creating a symbiotic loop.

The Data Science Periodic Table

Developed as a didactic tool, this table organizes data science concepts:

• Elements: Each cell denotes a specific element, identified by a two-letter symbol (e.g., ET for Extract, Transform, Load).
• Rows (Maturity of Data):
  • Row 1 (Raw Data): Includes ET and CD (Data Cleansing), dealing with data directly from its source.
  • Row 2 (Prepared Data): Features DI (Data Ingest) and EN (Data Encoding).
  • Row 3 (Model Data): Encompasses activities like Clustering and Aggregation.
  • Row 4 (Validated Insight): Focuses on GO (Data Governance) and auditing processes.
• Groups (Columns): Categorize elements by function: Acquisition, Preparation, Modeling, Generation, and Evaluation.
• Quantum Addendum: Contains advanced elements like Quantum Encoding and Quantum Metrics, which extend beyond classical data science.

The AI Periodic Table

Martin Keen’s AI periodic table is structured as follows:

• Groups (Columns):
  • Reactive: Agents that process input and act upon it.
  • Retrieval: The memory component of AI, involving EM (Embeddings) to convert information into numerical representations and VX (Vector Database) for storage.
  • Orchestration: Manages the coordination of various AI components, exemplified by RAG (Retrieval Augmented Generation).
  • Validation: Ensures the integrity and fairness of the system, including Guardrails and Red Teaming.
  • Models: Comprises LLMs (Large Language Models) and advanced Thinking Models.
• Rows:
  • Primitives: Fundamental building blocks: prompt, embed, and LLM.
  • Composition: Elements typically found in production environments, such as function calling.
  • Deployment: Refers to frameworks for deploying AI systems.
  • Emerging: New and rapidly developing concepts like multi-agent systems.

Building an AI Application: The Document Q&A Example

The speakers illustrate the practical integration of elements from both tables by constructing a document Q&A system for a finance analyst. The primary goals are to prevent AI hallucinations, ensure answers are grounded in company documents, and maintain strict access permissions.

The Linear Pipeline

This initial system represents a straightforward “reaction” where data progresses sequentially to generate an answer.

Data Science Elements (Pre-inference Data Preparation):

1. ET (Extract, Transform, Load): Gathers organizational policies (HR, generative AI) from disparate sources (SharePoint, Confluence, file shares, wikis) into a centralized location.
2. DI (Data Ingest): Ensures the pipeline is continuously updated with the latest policy revisions, preventing the use of outdated information.
3. CD (Data Cleansing): Removes extraneous content from documents (e.g., letterheads, footers, watermarks, OCR errors) from PDFs.
4. ST (Structured Data): Chunks each document by its sections and enriches it with metadata (e.g., department, effective date, sensitivity), making the data semantically queryable.
5. EN (Data Encoding): Converts categorical metadata (department, language) into a structured format suitable for system filtering, a process analogous to AI embeddings.
6. GO (Data Governance): Enforces access permissions and maintains audit trails, guaranteeing that the analyst can only retrieve authorized documents.

AI Elements (Runtime Execution): 7. EM (Embeddings): Transforms each cleaned document chunk into a numerical vector, facilitating semantic search by ensuring similar passages yield similar vectors. 8. VX (Vector Database): Stores these generated vectors, enabling rapid and efficient semantic retrieval at runtime. 9. RG (Retrieval Augmented Generation): Upon receiving a user query, it is embedded, the VX (Vector Database) is queried for the most relevant chunks, and GO (Data Governance) filters these results based on the analyst’s permissions. 10. PR (Prompt): The permitted, retrieved chunks are integrated alongside the analyst’s original question into a prompt template, providing essential “grounding” for the Large Language Model. 11. LG (Large Language Model): Generates an answer strictly based on the provided and grounded chunks, mitigating the risk of hallucination. 12. GR (Guardrails): Serves as a final output filter, verifying the accuracy of cited claims against source documents and redacting any Personally Identifiable Information (PII) before the answer is presented to the analyst.

From Linear Reaction to a Loop (Continuous Improvement)

To enable the system to learn and evolve autonomously, three additional elements are introduced to establish a feedback loop:

1. DR (Data Drift): Monitors the statistical distributions of data (e.g., query embeddings, user feedback rates). It triggers an alert when these distributions deviate significantly from a defined baseline, indicating potential performance degradation.
2. SD (Synthetic Data): When DR flags a drift, AI is employed to generate synthetic Q&A pairs specifically addressing the identified failing patterns.
3. FT (Fine-tune): These synthetic Q&A pairs are then used as training data for the embedding model (EM). This process teaches the model to accurately map specific questions to their corresponding relevant chunks within the vector space. The improved embedding model then updates the EM element, completing the loop.

This continuous feedback mechanism ensures the system learns from its operational experiences, leading to progressively better and more accurate answers for users without requiring manual intervention. This underscores the indispensable role of robust data science practices in underpinning successful real-world AI applications.

• Tags:
• Synthetic Data
• Embeddings
• RAG
• LLMs
• Feature
• Vector Database
• AI Architecture
• Data Governance
• AI Periodic Table
• Data Science Periodic Table
• Data Drift
• Document Q&A
• Machine Learning Pipeline
• Data Preparation